以深度学习为代表的人工智能技术凭借其高效性、通用性和可迁移性，在土木工程应用中有望弥补有限元等传统数值方法的局限性，近年来已经成为结构计算领域的重要研究方向。本论文基于深度学习技术，研究结构体系的智能计算方法，开发结构体系的智能分析模型，取得的主要研究成果如下：（1）建立了通用数据接口和基准数据集StructSet。针对现有的数字化方法的不足，提出了一种通用的结构体系图数据表示方法，基于该数据结构建立了结构体系数据集StructSet，补充了该领域大规模数据集的短缺。（2）开发了适用于结构体系弹性分析的智能计算模型StructGNN-E。分析了弹性计算的特点，提出了理论辅助驱动的模型训练方法，并基于图数据结构和图神经网络建立了模型架构。数值试验和消融试验验证了模型的有效性和准确性。（3）提出了结构体系的图表征学习模型和与训练方法。为了提高模型的泛化能力，提出将结构表征模型框架StructGRL，并设计了对比学习的预训练方法。将预训练的 StructGRL用于下游任务，验证了预训练方法的有效性，为进一步开发弹塑性分析模型奠定基础。（4）开发了适用于结构体系弹塑性分析的智能计算模型StructGNN-N。为了处理序列问题，引入基于序列到序列架构和注意力机制的Structformer网络，与图神经网络GINE组成StructGNN-N的基本框架，并提出基于全局数据的训练方法。结合预训练模型StructGRL开展了结构体系的地震时程分析，分别讨论模型对单个结构体系进行弹塑性分析的计算能力和推广到多个结构体系的泛化能力。（5）开发了适用于数字孪生模拟的智能推演模型StructGNN-D。讨论了利用深度学习方法实现工程结构的数字孪生模拟的技术路线，提出了由结构体系的部分响应推演全局响应的智能推演模型，并设计了基于部分数据的训练方法。最后通过结构试验数据检验了模型的有效性。本论文的课题研究由国家自然科学基金创新研究群体项目 (52121005) 与科学探索奖提供资助与支持。
Recent years have witnessed a growing interest in applying artificial intelligence (AI) techniques, represented by deep learning, to civil engineering, aiming to overcome the limitations of traditional numerical methods. This paper explores intelligent computational methods and models for structural systems based on deep learning techniques. The main research works and results are as follows:(1) Structural data representation scheme: An analysis of existing research reveals the deficiencies in digitalizing structural systems. Consequently, a universal graph data representation scheme is proposed for the structural systems, and the benchmark dataset StructSet is established to addresses the scarcity of large-scale datasets in this field.(2) Intelligent computational model for elastic analysis: StructGNN-E is developed for elastic analysis of structural systems. By analyzing the characteristics of elastic equations, a theory-informed training method is proposed. Based on graph data and graph neural networks, the model architecture is established and validated through numerical experiments and ablation studies.(3) Graph representation learning model: To enhance the generalization ability, a graph representation learning framework, StructGRL, is proposed to further transform structural system information into abstract feature embeddings. A contrastive pre-training method is designed and validated by downstream tasks, laying the foundation for further development of elastoplastic analysis models.(4) Intelligent computational model for nonlinear analysis: StructGNN-N is developed for nonlinear analysis of structural systems. To address sequence data, a model framework combining sequence-to-sequence architecture and attention mechanisms is introduced. Collaborating with the pre-trained model StructGRL, seismic analysis of structural systems is conducted to validate the model‘s capability to compute efficiently and generalize to different structural systems.(5) Intelligent prediction model for Digital Twin simulation: An intelligent prediction model called StructGNN-D and corresponding training methods is developed for digital twin simulation of engineering structures. The methods are validated through experimental data.This dissertation is sponsored by National Natural Science Foundation of China Program (Grant No. 52121005) and the China XPLORER PRIZE.
